Search and filters

import { retrieveDocuments } from "@anvia/core/vector-store";
import { vectorFilter } from '@anvia/core/vector-store';
const results = await retrieveDocuments({
    store: store,
    model: embeddings,
    query: 'reset a password',
    topK: 10,
    filter: vectorFilter.or(vectorFilter.eq('tier', 'public'), vectorFilter.gt('priority', 3))
});

The adapter runs vector search first and applies the filter afterwards with core's matchesVectorFilter, matching eq, gt, lt, and, and or against each candidate's metadata. When a filter eliminates candidates, the store widens its candidate limit until topK results survive, the table runs out of rows, or minScore excludes the remainder.

Filters match against the metadata stored at ingestion, so give filtered documents metadata up front. Retrieval filters do not replace authorization.

Search scores are 1 - distance for cosine (the default) and -distance for L2 and dot distances; they are meaningful only for the selected distance and corpus.